Orwa Mahmoud — Senior AI & Full-Stack Engineer
Profile · Dubai
Ten years of shipped software, the last two spent on production LLM agents — agentic RAG, MCP tooling, real-time voice, multi-tenant SaaS.
For the last decade, Orwa Mahmoud has taken products from an empty repository to production: agents, APIs, dashboards, and the Kubernetes and CI/CD underneath them. Since 2024, he has specialised in production LLM agents on LangGraph — agentic RAG, MCP tooling, real-time voice, and multi-tenant SaaS.
He builds in two languages and for two worlds: enterprise platforms that thousands of people rely on every day, and open-source tools developers install with one command. The common thread is ownership — taking the whole journey, from the first architecture sketch to the observability dashboards that prove it works.
He cares about clean architecture — DDD, hexagonal, CQRS — real observability, and quality gates that hold: his open-source work ships with ~99% test coverage and zero linter suppressions. ■
A career timeline in shipped systems.
Web development for local businesses — learning where software meets real operations.
Full-stack delivery in Angular, Node.js, and Symfony — client platforms across many domains.
Led and grew a backend team, delivering enterprise-scale platforms and migrations.
Production SaaS and e-commerce — connecting product execution with scale and business impact.
Platform and AI engineering: production agents, identity, payments, and end-to-end delivery.
Still building where engineering, product thinking, and execution meet. ■
Open source · React · Design System Infrastructure
A headless React data table built once and rendered seamlessly across seven leading UI kits — with a single, predictable contract for every use case.
AdaptTable is a headless, UI-agnostic React data table engineered for flexibility at scale. Built as a ten-package monorepo, it ships a clean core and pluggable adapters for Mantine, MUI, Chakra, Ant Design, Radix, and Tailwind/shadcn—so teams can standardize on one table contract while delivering native look & feel across different products and brands.
A single data-source contract powers both client- and server-side data. AdaptTable handles the hard parts: URL-synced state, column management, responsive mobile cards, optional virtualization for large datasets, and first-class internationalization and RTL support. The result is consistency for users, velocity for teams, and fewer trade-offs in production.
“What you've built is genuinely impressive — thank you for building and sharing AdaptTable with the community.”
One table contract, rendered native in every UI world. ■
AI Front Desk · Multi-Tenant · English / Arabic
Sight is a multi-tenant AI front desk that answers on a tenant's behalf, grounded in its own knowledge base and designed to automate confidently while handing off gracefully when needed.
Sight blends retrieval and reasoning to deliver accurate, contextual answers for every tenant it serves. A hybrid RAG pipeline — vector search, BM25, and reciprocal-rank fusion — surfaces the most relevant facts from documents, policies, and FAQs.
When confidence is low, it escalates seamlessly to a human inbox, preserving context and transparency. The architecture uses DDD/CQRS with LangGraph — modular, observable, and tenant-isolated by design.
It runs across web, WhatsApp, and Telegram in both English and Arabic, adapting to each tenant's knowledge, tone, and workflows — a front desk that feels like part of your team.
“Automation earns trust
when it knows its limits.”
Confident automation — and the judgment to hand off. ■
Claude Code Plugin · Unattended Runs · MIT
nightshift is a Claude Code plugin for the hours you are asleep. You write the checklist; hooks keep the agent on site until every box is ticked, under rules you set and it cannot bend.
Agents stop early. Asked for eight things overnight, one will ship the four easy items and defer the hard ones to "a focused session" — which is the session it was already in. Telling it not to is a sentence competing with everything else in context.
So completion lives in a file instead. A Stop hook refuses to end the session
while any - [ ] remains, and re-states the contract each time it
blocks. The list is greppable, survives a crash, and a resumed session picks
it up where it fell.
The other night-killer is a question at 02:40 that waits until morning. During a shift the ask-the-user tool is denied outright: the decision is parked with a sensible default chosen and the work carries on, ready for review over coffee. And when the night kills the session outright — a crash, an API 500 — a watchman outside it revives the same conversation and carries it to morning.
“Completion lives in a file,
not in a phrase.”
You sleep. It works. The full story → ■
A production-minded stack spanning agents,
platforms, apps, infrastructure, and identity.
A stack chosen for what ships to production. ■
Let's build something that actually ships.
Open to senior AI, platform, and full-stack roles, consulting, and ambitious agentic-AI work. The fastest way to reach me is email or LinkedIn.
Fast Routes
End of Issue • Thanks for Reading